Full-duplex voice
Voice agents that can listen while speaking, handle interruptions, and respond without waiting for the other person to finish.
Voice AI
A voice chat experience that's truly interactive, intelligent and sociable.
Audissey builds full-duplex, end-to-end voice AI: agents that listen while they speak, interrupt, backchannel, and respond naturally in real time. We're developing our own models for low-latency speech generation, turn-taking, and natural conversational timing — the technology behind voice agents.
The problem
Voice AI has gotten good at sounding human. The next problem is behaving socially.
Today's voice AI can sound remarkably human and still feel awkward to talk to. It waits too long. It misses the moment to jump in. It never says "mm-hm." It treats conversation as a strict exchange of turns.
Natural speech is not the same thing as natural conversation. That gap is what we're working on.
What we're building
Natural turn-taking
Models for detecting when someone is finished speaking, when to interrupt, and when a conversational backchannel is appropriate.
End-to-end speech
Speech in, speech out, developed in house and optimized for low latency and natural conversational timing.
How it works
Two custom models, built to work together.
Audio generation
Speech generated directly by the model, not stitched together from a text-to-speech pipeline. Pitch, pace and emphasis carry through end to end, so it sounds like someone talking, not something reading a script.
Turn-taking
An ensemble of dense models collectively synthesize tone, pacing, pauses and the running transcript.
That's what decides when to jump in, when to hold back, and when a quiet "mm-hm" is the right response.
The dataset
Trained on a synthetically augmented dataset, curated for:
- Streamed text input, for a low time to first audio
- Graceful pausing on interruption
- Continued speech through backchannel cues, so a "mm-hm" doesn't reset the turn
Stay in the loop
Leave your email for updates on the research.